Skip to content
Walking Inside a DeepSDF Latent Space, in the Browser

Walking Inside a DeepSDF Latent Space, in the Browser

Click a dot in the UMAP atlas and a chair resolves from a 256-dim code. Drag a slider and watch it melt into a lamp, meshed live from a neural signed distance function at 64³ or 128³. DeepSDF latent spaces you can actually walk around in. Three.js, WebGL, no download.

Details

What it is

Most latent-space visualizations stop at a scatter plot. LatentSDF Explorer doesn't. Click any dot in the UMAP atlas and a chair or a plane or a lamp resolves out of a 256-dimensional code, rebuilt in your browser from a neural signed distance function. Pick two shapes, drag the t slider, and watch one melt into the other: a sofa growing armrests, a table sprouting a fourth leg, the surface reforming live off a marching-cubes pass you can crank from 64³ to 128³ when you want the geometry to stop lying to you.

Prefer to watch the model think? Flip it to point clouds. There's even a geodesic path mode, so the morph follows the curve of the learned manifold instead of bulldozing a straight line through it. Small touch, big difference in what you actually learn from the in-between shapes.

It's a coursework project, NTU's AI6131 3D Deep Learning, by Tanmay Nargas, Aditya Goel, and Sneha Balasubramoni, and it reaches past what the brief asked for. DeepSDF trains a network to answer one stubborn question, "how far am I from the surface," for every point in space; the shape lives implicitly in the weights plus a latent vector, never as an explicit mesh until you ask for one. The team trained those latent spaces on ShapeNet-derived data, then built a Three.js front end to actually walk around inside them, the clustering improvements and the interpolations both, which is the part most papers leave as a static figure and a promise.

It all runs in the browser, no download, no GPU envy required. So poke the atlas, scrub an interpolation, swap the resolution mid-morph, then open the repo to see how the SDF gets meshed on the fly. Worth a follow while you're there.

How it’s made

Technology
WebGL
Framework & Tool
Three.js
Domain
AI/ML
Interaction
Drag Interaction